Multi-step Ahead Time Series Prediction via Bagging Trees Based Neighborhood

Locality based prediction has shown excellent performance in difficult multi-step forecasting problems. However, either neighbor search or sparse coding is usually used to define locality, which requires considerable computations. In this paper, we are investigating a locality-based time series forecasting based on bagging trees defined neighborhood. The proposed method is based on finding all the training examples that share the same leaves in the bagging trees ensemble with the test vector. Then we calculate the predicted target value based on the training examples respective weights (how many times each training example share the same terminal node in each tree with the test example). Two types of bagging trees are investigated; (classification and regression). We test our proposed method on the monthly time series data from M3 competition dataset with horizons range (1 to 18). We compare our results with other machine learning and statistical forecasting techniques. The proposed model outperformed almost all other techniques especially in higher horizons.